Bridging Computational and Clinical Strategies to Improve Presurgical Identification of Epileptogenic Networks

About one third of epilepsy patients are drug-resistant. Resective surgery remains a key treatment option but depends critically on accurate identification of the seizure onset zone (SOZ), which is still guided mainly by subjective visual inspection of electrophysiological signals. Network-based metrics derived from intracranial EEG (iEEG) have recently shown promise for SOZ identification, but their interpretation has remained disconnected from standard clinical procedures, while the reported performance often hinges on the choice of machine learning classifiers and summary scores. We analyzed interictal stereotactic EEG (sEEG) recordings from 20 patients undergoing presurgical evaluation, including cortical and subcortical implantations, with clinical mapping via electrical stimulation. We constructed patient-specific dynamic network models and compared the values of four corresponding metrics of network vulnerability (outgoing fragility, incoming fragility, source influence, sink connectivity) that were previously proposed as promising SOZ markers, with stimulation-evoked discharges. We also simulated virtual thermocoagulation by removing the clinically coagulated nodes and testing whether the resulting network changes went beyond pure network size reduction. The network metrics correlated with epileptiform discharges evoked by 50 Hz intracranial stimulation, directly linking model-based fragility with interictal epileptiform discharges evoked in clinical stimulation mapping. Using virtual thermocoagulation, we showed that the network models can predict the consequences of lesioning, capturing both local and global effects depending on individual network architecture. Across patients, the network metrics consistently distinguished SOZ from non-SOZ contacts and yielded stable conclusions across time, conditions and perturbation properties, supporting their reliability. Together, these findings show that iEEG-based network models provide clinically meaningful and interpretable markers of brain responsiveness to electrical stimulation, and can be used to predict the consequences of virtual resections. By relying only on interictal recordings, they avoid the clinical and technical challenges of capturing seizures and instead offer a personalized framework that complements presurgical mapping and guides surgical planning in drug-resistant epilepsy.